Patchwork Harness v0.4.1

A model refresh. Every model below was verified by a real call, and every price comes from the vendor's pricing page.

Models

• Gemini 4 Argon (gemini-4-argon) is in the catalog. It costs $2/$10 through 30 Nov 2026, then $4/$20, and the ledger bills the new price from 1 Dec. It takes 1M tokens in and 32,768 out.
  • It leads the Google lane of security_reviewer and is the Google pick in reviewer, ahead of Gemini 3.1 Pro.
  • On the reviewer calibration suite (once per case) it caught 8/8 planted bugs and passed 4/4 good solutions, at $0.013 a review, with no API errors.
• Claude Sonnet 5.5 replaces Sonnet 5 as the bulk executor and in reviewer. On the same suite it gave the same verdicts as Sonnet 5 (8/8 and 4/4), at $0.023 a review against $0.043.
• Claude Haiku 5.5 (claude-haiku-5-5) is the planner fallback. It costs $0.10/$0.50 up to a 100k-token prompt and $0.50/$2.50 above, and the adapter applies the tier. In a planner head-to-head, the critic approved 5/12 of its first drafts against 1/12 for Haiku 4.5.

Removed

• Superseded models, each replaced by a newer model from the same vendor at the same or a lower price: claude-opus-5, claude-sonnet-5, claude-haiku-4-5, gpt-5.6-sol, gpt-5.6-luna, gpt-5.5, gpt-5.5-pro, gpt-5.4-mini, o4-mini, o3-pro, gemini-3.7-flash and grok-4.6. An old id named explicitly still prices from the adapter's table.
• The gpt-6 and gpt-6-pro placeholders. They returned 404 on every probe since 3 Sept and never appeared on OpenAI's pricing page. OpenAI ships GPT-6 as Astra, Sol and Luna.

Fixed

• Gemini 4 got no thinking headroom, because only gemini-3* ids were treated as thinking models. Both now get headroom, clamped to each model's output cap, so a large request can't exceed Argon's 32,768.
• Haiku 5.5 had no price, so its calls would have gone on the ledger at $0. Haiku 5.x and Sonnet 5.5 also rejected temperature with a 400; it is no longer sent to them.
• Gemini 3.8 Flash's promotional price ends on 1 Jan 2027. The ledger bills $1.50/$7.50 from then.